Step5 Preview launches with 600B params, 27B active, 1M context — top-2 open model at 12.5% of Opus 5 cost
量子位 · wechat · 2026-09-21
StepFun quietly released Step5 Preview: a MoE flagship with 600B total / 27B active parameters and 1M context, scoring 44 on Artificial Analysis (global open-source top-2), priced at $1/M input and $2.7/M output — about 12.5% of Opus 5's per-task cost, sitting on the performance-cost Pareto frontier. Hands-on tests (Blender agent building a 'Backrooms' video, a LEGO Racers-style game, a parkour game, a writing site) show real agentic ability with minor bugs fixed in 2-3 rounds. Technically: a 92-layer 'Narrow but Deep' transformer for multi-hop reasoning, SparseMoE/HybridSparse/SparseGQA plus kernel-level optimization for long contexts, and Long-horizon RL with Context Compaction to sustain dozens of tool-call loops. The piece notes top models are now within 25 Elo of each other, pushing vendors toward differentiated strengths.
Related event: StepFun Launches Step 5 Preview: 600B MoE Flagship, Open-Weights on Oct 15(17 posts)→
More from coding & agent
- Ora's Ranker 3.0 grades whether agents can actually use your site; scores drop 40+ points — yenkel · 2026-11-12
- Gemini API Managed Agents now support self-hosted sandboxes, incl. Vercel — steren · 2026-11-12
- Rayline launches ARC, a per-turn model router claiming 90% cost cuts — testingcatalog · 2026-11-12
- Karpathy's llm-wiki: a pattern for building personal knowledge bases with LLM agents — stochasticchasm · 2026-11-12
- Agent self-improvement with frozen weights: failure traces add just 4pp, blind search gets 60% — burny_tech · 2026-11-12
- Microsoft on agentic schedulers: LLM proposes, deterministic code validates — WirelessLife · 2026-11-12